Jack Dorsey’s Block made an AI model that talks to you like an awkward new

Jack Dorsey’s Block launched Dessa, an artificial intelligence model built to communicate directly through spoken dialogue. Trained entirely from scratch on human speech recordings, the model operates without processing text transcripts. According to The Verge's report, the company demonstrated the system playing guessing games, discussing cinema, and responding with emotional inflection. The model represents a clear departure from traditional voice assistants that transcribe spoken audio into text tokens before synthesizing responses.

This report examines Dessa's speech-to-speech architecture, local hardware requirements, and early safety evaluations. The analysis is for software engineers, machine learning researchers, and system architects evaluating on-device speech interfaces. It covers demonstration dialogues, early research access limits, and conversational oddities observed during testing.

Jack Dorsey’s: the announcement

Block published release details and unedited recorded dialogues to show how the model manages spontaneous verbal exchanges. Project lead Ragavan Thurairatnam shared a demonstration of the system participating in an open conversational exchange with a human speaker. In that interaction, the model maintained an organic conversational rhythm, although it stumbled unexpectedly by mistakenly identifying the human speaker as someone named Ronaldo. Additional recordings on the project website showed the model playing an interactive guessing game conducted purely over real-time voice without written prompts.

Demonstration recordings on the website included conversations exploring cinema and emotional role-play scenarios testing dialogue limits. In one recorded exchange, a user discussed the films of director Christopher Nolan, prompting the system to branch into an unprompted tangent about director Bryan Singer and his Superman films. Another demonstration featured a human telling the model about an impromptu trip to Paris, which led Dessa to respond with an inflection sounding like teary gratitude. Block also published a comparison dialogue between OpenAI's GPT-Live-1 voice model and Dessa, where Dessa came across as more natural.

What actually changed with Jack Dorsey’s

The underlying architecture operates as a true speech-to-speech neural network rather than a conventional pipeline chaining speech recognition, text analysis, and audio synthesis. Block stated that every piece of information within the model came directly from human speech recordings. The engineering team noted that human babies learn to talk without ever seeing written words, prompting the researchers to see if an artificial intelligence model could do the same. Although the team acknowledged that speech-only training proved very hard, the resulting model directly preserves conversational cadence, tone, and timing.

Direct audio processing targets fundamental weaknesses in existing voice systems that convert spoken input into text. The developers argued that contemporary voice tools waste the natural power of speech by flattening expressive words into static text, which cannot match human conversational timing. Most importantly, current artificial intelligence remains deaf to emotional cues in human voices. The team stressed that a user in severe physical pain could speak to a voice assistant while the system remains unaware of the distress. Block positioned Dessa as a step toward resolving this safety vulnerability.

Who should care about Jack Dorsey’s

Artificial intelligence safety researchers and engineers developing voice interfaces should examine the model for insights into emotional audio processing. Block cited three explicit reasons for developing the project: curiosity, better mutual understanding between humans and artificial intelligence, and safety. Traditional voice models fail to interpret vocal panic or tone variations, creating hazards in high-stakes environments where vocal distress indicates an urgent crisis. By studying how a neural network processes acoustic tone directly, researchers can evaluate whether speech-native architectures offer more reliable comprehension when interacting with stressed human speakers.

Hardware specialists and developers focused on local edge deployment will also find the technical footprint significant. Because the model avoids external cloud dependencies during execution, it illustrates how high-fidelity speech synthesis can run locally without streaming private voice recordings across remote networks. The release also offers valuable data for engineers evaluating how models handle conversational interruptions, awkward pauses, and conversational missteps. Teams designing interactive voice systems can examine how removing text tokenization affects dialogue pacing, hallucination patterns, and odd identity mix-ups during real-time human conversation.

How to try Jack Dorsey’s

Direct access to the model remains closely guarded following the initial project announcement. Block stated that it will grant early access exclusively to a small group of artificial intelligence safety researchers to test the system and assess behavioral risks. The company has not provided a broad public beta release, commercial pricing structure, or general developer application programming interface. For now, interested users can listen to recorded voice samples, guessing game trials, and multi-model comparison audio clips published on the official Dessa website.

Users who eventually receive deployment builds will need high-end local desktop hardware to execute the system. According to the development team, the speech-to-speech network can run locally on an Apple MacBook Pro equipped with an M4 Max processor or better hardware. This on-device execution model ensures that speech recognition, reasoning, and synthetic voice generation operate directly on local silicon without sending audio packets to third-party cloud infrastructure. The team did not list support for older hardware platforms or less capable processor configurations.

What to watch after Jack Dorsey’s

Future milestones will depend on how effectively Block refines the model to eliminate conversational hallucinations and awkward dialogue tangents. Test recordings highlighted visible conversational quirks, such as the system confusing a human interlocutor with Ronaldo and abruptly shifting a conversation about Christopher Nolan to Bryan Singer. Observers will watch whether subsequent checkpoints preserve the emotional inflection heard in the Paris travel exchange while reducing erratic topic shifts. Successfully eliminating these factual and contextual errors without flattening vocal tone will test the viability of text-free voice training.

The broader technology sector will watch whether competing laboratories shift toward speech-to-speech training methodologies. Because Dessa displayed more natural back-and-forth cadence than OpenAI's GPT-Live-1 during demonstration recordings, competing developers may reevaluate whether text-based pipelines are suitable for conversational products. Technical reports from the initial group of safety researchers will indicate whether speech-native training reliably prevents models from missing urgent vocal distress. Observers will also track whether on-device processing on Apple M4 Max hardware can support sustained interactive dialogue without external cloud assistance.